i7 920 or 965

[ctrlbrk]

I currently have an E6850 3.0ghz overclocked to 3.51. I run large optimization jobs on my system which peg it at 100% cpu for hours, days, sometimes a week solid on end.

I very much need a faster CPU. But, money is an object (not unlimited budget) so I know I am buying a i7 but I don't know which one --- 920 or 965.

Sure, I've read the reviews on the 920 overclocking to 3.2, 3.5, even 4.0. But, lets remember this machine will be at full load for days sometimes, this has to be rock-solid stable, so I wouldn't feel comfortable based on the voltages I've been reading about taking it much past 3.2 or 3.5.

So, now enter the 965. The question is, how far can the 965 go on 'normal' conditions again making it 100% rock solid stable? I feel like it would have to be at least 4ghz before I would consider it worth the extra money. But I am not sure it will get there.

Thoughts?

 

[Idontcare]

If you are truly clockspeed limited, and time is money, then you ought to consider getting the 965 (just to make sure you are getting all the GHz you can out of your i7) and combine it with a vaporphase cooling system so you can clock it to the 5GHz region for 24x7 usage.

I was in similiar situation (time was money, needed GHz, etc) a couple years ago so I went with a QX6700 (2.66Ghz unlocked Kentsfield quad) and vaporphase rig to run 4GHz 24x7 for a year. It was spendy, but I got what I paid for and if you wanted a 4GHz quad system in 2006 it was the only way to get their.

If you want a 5GHz 24x7 stable i7 system then you want the 965 and a vapochill lightspeed cooler (wait for the LGA1366 clamshell though).

Also consider eliminating performance bottlenecks that come from spindle-drives by installing a nice 2GB cache Areca controller card with 3 or 4 raid-0 Intel SSD's.

 

[ctrlbrk]

IDC,

My drives are fine (RAID 0, around 185MB/sec) -- I am CPU limited.

Let me re-iterate that while time IS money, I in fact DO have a budget so I am willing to throw down for new memory, new mainboard, and new CPU... I am having a hard time determining if it is worth an extra $700 bucks for the 965.

Assuming a top of the line AIR cooler, will the 920 oc to 3.2 on normal voltage good for 24x7 stable operation at full load?

If so, now how much further would the 965 go on the same top of the line AIR cooler with normal voltage for 24x7 full load use?

That is the answer I don't have. A zillion benchmarks on the 920 oc but I have read very little on the 965 oc at STABLE voltages without liquid cooling, which I don't want to get into right now.

Thanks!

 

[nyker96]

I think 920 can do closer to 4 if I read some preliminaries correctly. No matter how far 965 can go say 5 ghz, it cannot give you a good performance/price ratio, that's definite. But for people money is no object then only pure performance matters for them and 965 of course will win.

 

[Markfw]

Originally posted by: ctrlbrk
IDC,

My drives are fine (RAID 0, around 185MB/sec) -- I am CPU limited.

Let me re-iterate that while time IS money, I in fact DO have a budget so I am willing to throw down for new memory, new mainboard, and new CPU... I am having a hard time determining if it is worth an extra $700 bucks for the 965.

Assuming a top of the line AIR cooler, will the 920 oc to 3.2 on normal voltage good for 24x7 stable operation at full load?

If so, now how much further would the 965 go on the same top of the line AIR cooler with normal voltage for 24x7 full load use?

That is the answer I don't have. A zillion benchmarks on the 920 oc but I have read very little on the 965 oc at STABLE voltages without liquid cooling, which I don't want to get into right now.

Thanks!

Isn't the 6850 just a dual-core ? I simple Q9550 would more than double your capacity, and might even work on the same motherboard. The I7 would cost you at least $1000 to upgrade, and at best would do 40% better than the Q9550, and you can;t even buy them yet anywhere that I have seen.

 

[Idontcare]

Originally posted by: ctrlbrk
IDC,

My drives are fine (RAID 0, around 185MB/sec) -- I am CPU limited.

Let me re-iterate that while time IS money, I in fact DO have a budget so I am willing to throw down for new memory, new mainboard, and new CPU... I am having a hard time determining if it is worth an extra $700 bucks for the 965.

Assuming a top of the line AIR cooler, will the 920 oc to 3.2 on normal voltage good for 24x7 stable operation at full load?

If so, now how much further would the 965 go on the same top of the line AIR cooler with normal voltage for 24x7 full load use?

That is the answer I don't have. A zillion benchmarks on the 920 oc but I have read very little on the 965 oc at STABLE voltages without liquid cooling, which I don't want to get into right now.

Thanks!

SSD's for latency improvement, not bandwidth. But if you know your apps well enough to rule it out then you know all you already need to know, which is a great place to be.

For my apps I was writing/reading a ton of small files, turned out the IOP's for 4k random file read/writes was a good metric for my rate-limiting performance. SSD's weren't around at that time but Gigabyte Irams were, so I used them to much satisfaction.

I have no i7 hardware or data to answer your questions, hopefully others will.

 

[ctrlbrk]

Originally posted by: Markfw900
Isn't the 6850 just a dual-core ? I simple Q9550 would more than double your capacity, and might even work on the same motherboard. The I7 would cost you at least $1000 to upgrade, and at best would do 40% better than the Q9550, and you can;t even buy them yet anywhere that I have seen.

Yes the 6850 is a dual core. The optimizations I run are very math intensive, floating point, integer, etc. Based on what I've read, i7 offers a significant advantage over C2D/C2Q in this specific arena.

I could throw in a C2Q into my current system but if I can get another 50% from i7 I will take it. Also, the optimizations I run are single threaded. I am hoping HT would help them out. I already run a VM instance to maximize the second core, and with a i7 (or C2Q) I know more VM's would be necessary to fully utilize every core, but the more VM's I do the more overhead I get with my optimizations. I prefer the fastest "core" possible to multiple slower cores.

Thanks for all the input.

 

[ctrlbrk]

Originally posted by: Idontcare
SSD's for latency improvement, not bandwidth. But if you know your apps well enough to rule it out then you know all you already need to know, which is a great place to be.

Thanks IDC. Yes actually I am very well versed in raid, I was VP Eng/Ops for many years at a leading raid manufacturer. I do want some X25's but right now are out of my budget and not necessary to my app, just to my ego.

Thanks again.

 

[Zap]

Originally posted by: ctrlbrk
Also, the optimizations I run are single threaded. I am hoping HT would help them out.

Why would HT help out single threaded apps that extra "real" cores wouldn't?

Originally posted by: ctrlbrk
I prefer the fastest "core" possible to multiple slower cores.

If so, then how about looking for a sweet overclocking Wolfdale? Probably wouldn't be too difficult to find an E8600 that can do 4.5GHz with good cooling. Check out this thread. First post showed 4.5GHz. Later on, two people got 5GHz. That's only the first page.

Here's a spreadsheet showing a bunch of results that someone compiled. The MINIMUM result for an E8600 was 4.0GHz, and some of those overclocks are under 1.2v.

Woot! Second page someone got just about 5.8GHz, but on a whopping 1.824v! :Q

Anyways, I'm not sure that Core i7 will regularly be capable of as fast single threaded overclocked performance anytime soon.

 

[aigomorla]

Dammit, the E8600 is sooooo not fair....

The only Quadcore counterpart to that chip is the Uber rare X5470.

And in the right board... it can wreak some uber damage. Stupid E0 stepping.

 

[ctrlbrk]

Zap,

You are correct on the HT. Duh. I have been starring at monitors looking at code for about 3 days now and wasn't thinking, clearly.

As for the Wolfdale. Interesting idea. My E6850 maxed at around 3.51Ghz (where it is now) for my 'comfort level' on 24x7 stability. Do you really think taking an E8600 to 4, or 4.5ghz would make a dramatic difference?

I was under the impression that Nehalem's number crunching capabilities, even clock for clock, were far better than the C2D or C2Q's. So even if Nahelem was only running at the same 4ghz clock, it would still stomp all over the Wolfdale.

Am I wrong?

Thank you again for your insight.

 

[aigomorla]

Originally posted by: ctrlbrk
I was under the impression that Nehalem's number crunching capabilities, even clock for clock, were far better than the C2D or C2Q's. So even if Nahelem was only running at the same 4ghz clock, it would still stomp all over the Wolfdale.

Am I wrong?

Thank you again for your insight.

no they showed clock per clock on a single threaded application yorkfield even was slightly faster. Tiny bit...

Honestly there is no real "advantage" to neha, unless your running high end videocards, and you need that massive memory bandwith.

However once better multithreaded applications come out, i7 with HT pretty much dominates across the board.

And i bet the transition from dualcore -> quadcore will be a lot faster then single core -> dual core.

 

[ctrlbrk]

Originally posted by: aigomorla
no they showed clock per clock on a single threaded application yorkfield even was slightly faster. Tiny bit...

Thank you for this. Do you have a link or can tell me which site? I must have not seen this particular comparo. I was looking at drystone and wetstone and crap like that trying to get idea for floating point speed since my app is mainly doing math.

 

[Dadofamunky]

I think a Wolfdale would be a great stopgap for you and a good cost-benefit choice for six months or so. But that thread that Aigo posted shows many of them are cranking the volts up beyond what is safe for those chips. 1.3625 is the 'official (Intel)' limit for Wolfdales. The only way to get away with that is with major water or phase-change, which it doesn't sound like you're interested in (neither am I). The rig in my sig is Linpack stable at 4.12 GHz, 1.35V set in BIOS. That's as high as I'm willing to take it. At the same clock speed, the Wolfies have about, what, a 15-20% performance boost over the Conroes. I'm very skeptical of anyone claiming 4.4-4.5 GHz OCs on stock voltage or near there. A good mobo sure doesn't hurt, though!

***EDIT*** some people ARE getting MONSTER OCs with P45/P48 mobos, however, and keeping the volts safe. My mobo is kinda weak... but uATX beggars can't be choosers.

 

[Tempered81]

Originally posted by: ctrlbrk
I prefer the fastest "core" possible to multiple slower cores.

Thanks for all the input.

4.5Ghz E8600

 

[Tempered81]

Originally posted by: ctrlbrk

I was under the impression that Nehalem's number crunching capabilities, even clock for clock, were far better than the C2D or C2Q's. So even if Nahelem was only running at the same 4ghz clock, it would still stomp all over the Wolfdale.

it's about 25% than the penryns at the same clock (in certain situations like yours). 3ghz = 4ghz.

Probably a 3.6ghz i7 920 would equal a ~4.5ish ghz e8600

 

[BTRY B 529th FA BN]

I would go with the middle of the road i7

 

[Idontcare]

Originally posted by: ctrlbrk

Originally posted by: aigomorla
no they showed clock per clock on a single threaded application yorkfield even was slightly faster. Tiny bit...

Thank you for this. Do you have a link or can tell me which site? I must have not seen this particular comparo. I was looking at drystone and wetstone and crap like that trying to get idea for floating point speed since my app is mainly doing math.

Quick question about your optimizations...usually an optimization is an iterative process. While the application itself may be single-threaded you might still benefit from running multiple instances and running optimizations in parallel by setting each instance of the application to iterate over just a subset of the optimization space.

I use a single-threaded app, metatrader, and do backtest optimizations in this manner. On my quads I load four instances of the application and then I set each instance to analyze the backtest parameter space for just 1/4 of the total parameter combinations (each instance then analyzes a unique set of parameter conditions).

I get linear speedup this way, 4 cores means my optimization time is exactly 1/4 what it would be were I to run just a single instance of this single-threaded app. Just a suggestion, not all optimizations can be broken up and parsed this way though so it may be useless suggestion for you.

 

[Nemesis 1]

Next stepping normal O/C water 4.5 ghz.

 

[ctrlbrk]

Thanks all for the replies. Will post back here when I decide..